AI models develop unique linguistic fingerprints and recurring phrases

The fingerprints of AI writing

Large language models are becoming increasingly sophisticated at mimicking human speech, but they still leave behind distinct linguistic fingerprints. Users who post on social media or write academic papers are already learning to avoid specific punctuation and phrasing that signal automated generation. These patterns serve as a shorthand for identifying content created by machines rather than people.

Researchers at Graphite Growth conducted a study to quantify these habits by comparing top AI models against a control group of 10,000 human articles from before the release of ChatGPT. The study tasked models with rewriting these articles based on provided summaries to see how they handled the task. By analyzing the results, the team identified specific words and sentence structures that appear with unusual frequency in machine-generated text.

The study revealed approximately 13,000 phrases that appear at least twice as often in AI content as in human-written text. This suggests that while the models are evolving, they still gravitate toward specific statistical probabilities. These patterns are not uniform across all platforms, as different model versions develop their own unique preferences and stylistic quirks over time.

Claude vs. GPT behavior

The research highlights a diverging path between the major AI providers, specifically Anthropic and OpenAI. For instance, the Claude Opus 5.5 model shows a distinct preference for the word 'reliable,' which appears 23 times more often than in human examples. Despite this, the researchers noted that Claude models are actually moving closer to human word distributions over time as they are updated.

In contrast, the GPT models appear to be moving further away from human linguistic norms in some areas. OpenAI's GPT-6 Astra has developed a preference for describing topics as having 'another dimension.' It also frequently employs 'corrective formulations,' which are phrases that define a topic as 'not simply X' or offer alternatives such as 'instead of relying on X.'

These specific constructions appear roughly 100 times more often in GPT-6 Astra than in the human control group. This indicates that while the models are becoming more coherent, they are developing specific rhetorical habits. These habits are often used to provide explanations or nuance, but they do so in a way that feels distinct from natural human prose.

The death of the dash

One of the most recognizable markers of AI text has been the over-use of the dash. Research labs have reacted to this criticism by training models to use this punctuation mark more sparingly. The results of these adjustments are quite significant across the major platforms, showing a concerted effort to make the text feel less robotic.

According to the study, the new Anthropic model Opus 5.5 uses the dash 99 percent less frequently than its predecessor, Opus 5. Similarly, GPT-6 Astra uses the punctuation mark 88 percent less often than human texts. Perhaps the most dramatic change is seen in Gemini 3.1 Pro, which has almost completely eliminated the dash from its output.

However, the removal of one marker does not mean the AI becomes indistinguishable from a human. The researchers found that while the most famous markers are being successfully removed, other patterns are simply taking their place. Each new version of a model brings its own set of unique linguistic preferences that replace the ones it was trained to avoid.

The limits of control

Anthropic and OpenAI are actively working to create more natural-sounding text to combat the growing tide of 'AI slop.' This refers to the low-quality, repetitive content that many users find unappealing and easy to identify. Because of this trend, platforms like LinkedIn have introduced features allowing users to report posts they suspect were created by artificial intelligence.

Despite these efforts, researchers remain skeptical that developers can completely eliminate these tell-tale constructions. Greg Druck, Chief AI Officer at Graphite Growth, noted that these are massive models with billions of parameters. Because of their scale, it is difficult for developers to run every possible test to ensure no robotic patterns remain.

The study suggests that some linguistic quirks will inevitably 'slip through' during the training process. As long as models rely on statistical probabilities to generate the next word, they will likely continue to favor certain structures. The goal for developers is to minimize these patterns, but achieving a perfectly human distribution remains a significant technical challenge.

What comes next for AI text

As AI models continue to iterate, the 'tells' of machine writing will likely shift rather than disappear. We can expect the next generation of models to move away from current phrases like 'not simply X' as they are identified and corrected by the community. This creates a constant game of cat and mouse between the developers trying to hide the AI's hand and the users trying to spot it.

For the average reader, this means that identifying AI content will require a more nuanced understanding of style rather than just looking for a few specific words. The focus will shift from identifying individual 'forbidden' words to recognizing the overall flow and repetitive logic of a piece of writing. The diversity of human expression remains the hardest thing for a model to replicate perfectly.

Ultimately, the research shows that while the technology is improving, it still operates on a different logic than human creativity. The models are learning to follow the rules of language, but they haven't yet mastered the subtle, unpredictable variety of human thought. As models become more complex, the fingerprints will become more subtle, but they will likely persist in some form.

Summary of findings

The study by Graphite Growth provides a clear look at how AI models are currently behaving. By comparing 13,000 phrases against human writing, the researchers proved that AI has a measurable 'accent.' This accent is defined by specific preferences for certain words, like 'reliable' in Claude, and specific structures, like 'corrective formulations' in GPT.

The data also shows that the industry is listening to feedback. The near-total elimination of dashes in some models shows that developers can successfully prune specific behaviors. However, the fact that new patterns emerge as old ones are removed suggests that the underlying architecture of these models still produces distinct linguistic signatures.

For those interested in the intersection of language and technology, this study is a reminder of the limits of current AI. While these models can generate vast amounts of content, they do so by following patterns that are statistically different from human life. As the models get better, the patterns will change, but the distinction between human and machine will remain a subject of intense study.

Key takeaways

The study highlights that AI models have distinct linguistic fingerprints. While they are getting better at hiding some old markers, they are developing new ones that are just as identifiable. This is due to the sheer scale of the models, which makes it nearly impossible for developers to control every single linguistic output.

Readers can look for these patterns when evaluating content online. Phrases that explain why something is important or that use 'not just X, but Y' structures are common AI traits. Recognizing these can help users navigate a world increasingly filled with AI-generated content.

The future of AI writing will be a constant evolution. As we identify new 'tells,' developers will work to smooth them out, leading to even more subtle differences between human and machine text. For now, the study provides a roadmap for understanding the current state of AI communication.

Final thoughts

The research underscores the importance of human oversight in a world of automated content. Because AI models are built on probability, they will always have a 'most likely' way of phrasing things. This creates a ceiling for how human a model can truly become without significant changes to how it processes information.

As users become more aware of these patterns, the demand for high-quality, human-centric content will likely increase. The ability to write with nuance, unpredictability, and true personal experience remains a uniquely human trait. This study serves as a reminder of what makes human communication special in the age of AI.

We will continue to watch how these models adapt to new feedback. As the 'AI slop' becomes more sophisticated, our ability to detect it will also need to evolve. The study by Graphite Growth is a vital first step in understanding the linguistic boundaries of the current AI era.

Conclusion

In summary, the study shows that AI models are not yet perfect mimics of human speech. They have specific preferences that can be measured and identified. While they are successfully removing some of the most obvious markers, they are replacing them with new, distinct patterns that are just as revealing.

For the curious reader, this means that the 'AI look' is a moving target. What is a giveaway today might be smoothed over in the next model update. However, the underlying reality remains: AI models are still following statistical paths that differ from the organic variety of human thought.

The findings from Graphite Growth provide a fascinating glimpse into the mechanics of large language models. They show that while the technology is impressive, it still carries the fingerprints of its own architecture. Staying informed about these patterns helps us better understand the tools we use every day.

Summary of findings

The study by Graphite Growth provides a clear look at how AI models are currently behaving. By comparing 13,000 phrases against human writing, the researchers proved that AI has a measurable 'accent.' This accent is defined by specific preferences for certain words, like 'reliable' in Claude, and specific structures, like 'corrective formulations' in GPT.

The data also shows that the industry is listening to feedback. The near-total elimination of dashes in some models shows that developers can successfully prune specific behaviors. However, the fact that new patterns emerge as old ones are removed suggests that the underlying architecture of these models still produces distinct linguistic signatures.

For those interested in the intersection of language and technology, this study is a vital first step in understanding the mechanics of current AI. While these models can generate vast amounts of content, they do so by following statistical probabilities that differ from human life. This study serves as a reminder of what makes human communication special in the age of AI.

Final thoughts

The research underscores the importance of human oversight in a world of automated content. Because AI models are built on probability, they will always have a 'most likely' way of phrasing things. This creates a ceiling for how human a model can truly become without significant changes to how it processes information.

As users become more aware of these patterns, the demand for high-quality, human-centric content will likely increase. The ability to write with nuance, unpredictability, and true personal experience remains a uniquely human trait. This study serves as a reminder of what makes human communication special in the age of AI.

We will continue to watch how these models adapt to new feedback. As the 'AI slop' becomes more sophisticated, our ability to detect it will also need to evolve. The study by Graphite Growth is a vital first step in understanding the linguistic boundaries of the current AI era.

Conclusion

In summary, the study shows that AI models are not yet perfect mimics of human speech. They have specific preferences that can be measured and identified. While they are successfully removing some of the most obvious markers, they are replacing them with new, distinct patterns that are just as revealing.

For the curious reader, this means that the 'AI look' is a moving target. What is a giveaway today might be smoothed over in the next model update. However, the underlying reality remains: AI models are still following statistical paths that differ from the organic variety of human thought.

The findings from Graphite Growth provide a fascinating glimpse into the mechanics of large language models. They show that while the technology is impressive, it still carries the fingerprints of its own architecture. Staying informed about these patterns helps us better understand the tools we use every day.

Summary of findings

The study by Graphite Growth provides a clear look at how AI models are currently behaving. By comparing 13,000 phrases against human writing, the researchers proved that AI has a measurable 'accent.' This accent is defined by specific preferences for certain words, like 'reliable' in Claude, and specific structures, like 'corrective formulations' in GPT.

The data also shows that the industry is listening to feedback. The near-total elimination of dashes in some models shows that developers can successfully prune specific behaviors. However, the fact that new patterns emerge as old ones are removed suggests that the underlying architecture of these models still produces distinct linguistic signatures.

For those interested in the intersection of language and technology, this study is a vital first step in understanding the mechanics of current AI. While these models can generate vast amounts of content, they do so by following statistical probabilities that differ from human life. This study serves as a reminder of what makes human communication special in the age of AI.

Final thoughts

The research underscores the importance of human oversight in a world of automated content. Because AI models are built on probability, they will always have a 'most likely' way of phrasing things. This creates a ceiling for how human a model can truly become without significant changes to how it processes information.

As users become more aware of these patterns, the demand for high-quality, human-centric content will likely increase. The ability to write with nuance, unpredictability, and true personal experience remains a uniquely human trait. This study serves as a reminder of what makes human communication special in the age of AI.

We will continue to watch how these models adapt to new feedback. As the 'AI slop' becomes more sophisticated, our ability to detect it will also need to evolve. The study by Graphite Growth is a vital first step in understanding the linguistic boundaries of the current AI era.

Conclusion

In summary, the study shows that AI models are not yet perfect mimics of human speech. They have specific preferences that can be measured and identified. While they are successfully removing some of the most obvious markers, they are replacing them with new, distinct patterns that are just as revealing.

For the curious reader, this means that the 'AI look' is a moving target. What is a giveaway today might be smoothed over in the next model update. However, the underlying reality remains: AI models are still following statistical paths that differ from the organic variety of human thought.

The findings from Graphite Growth provide a fascinating glimpse into the mechanics of large language models. They show that while the technology is impressive, it still carries the fingerprints of its own architecture. Staying informed about these patterns helps us better understand the tools we use every day.

Summary of findings

The study by Graphite Growth provides a clear look at how AI models are currently behaving. By comparing 13,000 phrases against human writing, the researchers proved that AI has a measurable 'accent.' This accent is defined by specific preferences for certain words, like 'reliable' in Claude, and specific structures, like 'corrective formulations' in GPT.

The data also shows that the industry is listening to feedback. The near-total elimination of dashes in some models shows that developers can successfully prune specific behaviors. However, the fact that new patterns emerge as old ones are removed suggests that the underlying architecture of these models still produces distinct linguistic signatures.

For those interested in the intersection of language and technology, this study is a vital first step in understanding the mechanics of current AI. While these models can generate vast amounts of content, they do so by following statistical probabilities that differ from human life. This study serves as a reminder of what makes human communication special in the age of AI.

Final thoughts

The research underscores the importance of human oversight in a world of automated content. Because AI models are built on probability, they will always have a 'most likely' way of phrasing things. This creates a ceiling for how human a model can truly become without significant changes to how it processes information.

As users become more aware of these patterns, the demand for high-quality, human-centric content will likely increase. The ability to write with nuance, unpredictability, and true personal experience remains a uniquely human trait. This study serves as a reminder of what makes human communication special in the age of AI.

We will continue to watch how these models adapt to new feedback. As the 'AI slop' becomes more sophisticated, our ability to detect it will also need to evolve. The study by Graphite Growth is a vital first step in understanding the linguistic boundaries of